CLF3D: A Coarse-Labeling Framework to Facilitate 3D Object Detection in Point Clouds

Abstract

Tremendous scenarios have to be considered for autonomous driving, leading to extremely large amount of point cloud data which need to be labeled for model training. Manually labeling such data is labor-intensive and highly expensive. In this paper, we propose CLF3D, a simple and effective coarse-labeling framework designed to improve existing automated labeling methods by fully leveraging scene-specific information in the unlabeled data to significantly enhance detection accuracy. Specifically, CLF3D first utilizes a pre-trained model to generate initial pseudo-labels, which are subsequently refined using a two-stage filtering strategy in combination with an instance bank built from high-quality annotated instances. These refined pseudo-labels are then used to fine-tune the model, progressively improving its detection performance on unlabeled data. Through iterative refinement of pseudo-labels, the model parameters, and the instance bank, CLF3D continuously improves label quality and accuracy. Experimental results demonstrate that the proposed method improves the detection accuracy by up to 14% compared to the originally pre-trained model across datasets of various sizes. This means our approach can reduce 14% of the manual workload for labeling point cloud data in comparison to the existing methods.